Yanlong Yang
← All projects

Research2023 – 2025

RadarINV — Unsupervised Radar Point-Cloud Enhancement

First author · arXiv 2025

Framework: a diffusion prior trained on LiDAR BEV images, then posterior sampling guided by the radar forward model (range FFT → Doppler → beamforming).
Framework: a diffusion prior trained on LiDAR BEV images, then posterior sampling guided by the radar forward model (range FFT → Doppler → beamforming). RadarINV paper
sparse radar → LiDAR-like densitypriorlearned from arbitrary, unpaired LiDAR
Schematic

A small automotive radar array cannot resolve nearby targets in angle, so its point clouds are sparse and noisy. RadarINV enhances them toward LiDAR-like density using a diffusion prior learned from arbitrary, unpaired LiDAR data — removing the need for paired radar/LiDAR supervision.

The project began at Vanjee as radar super-resolution from raw radar signals and continued into my PhD.

Highlights

  • Unsupervised: no paired radar/LiDAR data
  • Diffusion prior from arbitrary LiDAR datasets
  • Array super-resolution as an inverse problem
arXivCode

Publication

2025Preprint

★Unsupervised Radar Point-Cloud Enhancement via Arbitrary LiDAR-Guided Diffusion Prior

Y. Yang, J. Liu, G. Luo, H. Li, E. Ahn, M. Rahimi Azghadi, T. Huang

arXiv:2505.09887

Enhances sparse, noisy radar point clouds toward LiDAR-like density with a diffusion prior learned from arbitrary, unpaired LiDAR data.

arXivCodeProject